The Future of Work in an Era of Agents: From Workslop to Work-for-Trust

If, at Anthropic, one of the frontier AI labs, a system combining tens of thousands of agents is already responsible for more than 25% of the work involved in developing new models, then the question is: does an organizational chart really reflect the reality of a company?

And that is without considering how difficult it had already become to read those charts accurately for more than a decade, knowing that we allow them to coexist with collaborative ecosystems based on agile methodologies, where professionals’ time is allocated to projects reviewed several times a year.

Now let us add the value layer of AI, which becomes the “alter-ego-worker” of more than two-thirds of your company’s employees—who also decide to use their preferred subscription, paid for out of their own pockets, when their monthly allocation of millions of corporate tokens runs out.

Do we really know which of the old corporate simplifications we use at work—organizational charts, professional directories on the intranet, job descriptions—still reflect the reality of our organizations?

Describing the workplace has been complex enough since the global upheaval of 2020. Hybrid working environments accelerated by digital technologies, multiple generations in the workplace, corporate agile practices, and the latest guest at the party: artificial intelligence “seeding” every kind of office work. Can we even begin to grasp what we have turned human collaboration into in the pursuit of business value?

The social contract through which I committed my time to a business endeavor has been redefined, without a corresponding regulatory framework that comprehensively governs this new relationship between the two parties.

WORKING IN AUGMENTED ENVIRONMENTS

Fred Kofman argued that your job is not what you do, but the goal you pursue [within the scope of your social contract with the business endeavor you are part of, I would add]. If your job is the goal—the desired end result agreed with your manager or internal client—what conditions define the scope of collaboration between companies and workers?

What framework and conditions govern the collaboration of this new augmented worker, who works alongside an extension of their body capable of generating more than 25% of the output, the desired end result of their efforts? Can we really express or agree on that commitment in units of time when I can find a combination of technologies that allows me to complete two weeks’ work in one hour with AI?

Last week, I spoke with an executive at an IBEX 35 company who admitted that his excitement about using AI had led him to give up his vacation this year, entirely by personal choice, driven by anxiety over how much work he could get ahead on with AI. I think something has broken.

Creating sustainable working environments requires a thorough reexamination of the ecology of work in the AI era, along with some initial principles that can lay the foundations for new discussions—before we arrive at the widespread realization that what we used to call talent in business has broken down.

We need to recognize that work in the AI era is an organizational architecture. An operating system through which an organization distributes intelligence, execution, autonomy, and responsibility among people, teams, and now machines. Failing to act accordingly means accepting that our definition of work “has already broken down,” while choosing to look the other way.

Just as housing markets in some countries have broken down through a lack of foresight, the labor market has transformed before we even grasped the change that had been approaching since late November 2022. And this latest change is global.

• A culture of intrapreneurs is no longer the same as a finely tuned operating model built around intrapreneurs.

• Hybrid work since 2020 has revealed purposeful presence as a new requirement.

• We already know that agile methodologies can exist within rigid, unadaptable organizations run by leaders stuck in the past.

• Diversity is no longer a token gesture toward representation; it is a pillar of governance if we want genuinely balanced environments for human–machine collaboration.

• And AI is no longer merely an automation solution; it is an essential ingredient in augmented working environments and augmented professionals. It is the lifeblood connecting this reality of augmented teams.

If we can observe greater individual autonomy at scale and through automated means; if the degree of full automation in certain tasks—research, more than 90% according to Deloitte—means that individual value creation resides primarily in the layer of human supervision; if the job must be replaced by a “Task → Workflow → Roles → Team → Organization” perspective, and hierarchies by a view of human–agent networks; if understanding the acceleration of AI use is no longer about measuring tool adoption but about observing its transformational impact from a bird’s-eye view; if humans are a thousand times more valuable when they debate, generate ideas, and discuss coexistence between humans and machines than when they simply execute, should we not redefine the concept of a job and the relationship between professionals and business endeavors—I deliberately exclude the employee–company framing—and also redefine how trust is built in a new ecosystem of collaboration between humans and machines?

If a professional prefers to conceal the fact that they can now complete weeks of their previous work in an hour, if an employee hesitates over whether to report to their manager that a colleague’s efficiency comes from personal AI rather than corporate AI, has trust not already broken down?

What if a workflow were more than a set of tasks constrained by people’s capabilities and time? We cannot keep thinking in terms of traditional humans. We need to look through the lens of augmented humans.

Let us remember that work is hybrid in three senses: place (in person/remote), time (synchronous/asynchronous), and the nature of the performer (human/agent). This combination of dimensions creates workflows that a traditional mindset cannot imagine. I can myself create work queues or batch jobs to run during my resting hours. In person or remote is a distinction we need to revisit, since it amounts merely to policies for controlling people or coordinating teams with traditional, unaugmented mindsets, and these choices create productivity bottlenecks. Finally, most of any human being’s productivity no longer comes from the person themselves, but from their ability to design asynchronous work for agents to carry out while they rest.

I do not think we are aware of the new landscape emerging simply from combining synchronous and asynchronous work. If we were, our first response would be to feel overwhelmed, and our second would be to begin reflecting on this new definition of work, which disrupts how work is organized and demands measures to safeguard the mental health of the human members of the team.

I believe we are on the brink of a breakdown of trust in the very concept of work.

MANAGING TRUST IN AGENTIC ENVIRONMENTS

If trust has broken down, can we restore it with the help of that digital companion we call Augmented Intelligence?

My proposal is to replace environments of execution with environments of exploration and learning, and performance networks with learning teams. Work-for-Trust as a new architecture of work built around continuous learning, with the continuous creation of trust in human–machine interactions as its North Star.

In her definition of the Learning Zone, Amy Edmondson explained that when our attitude at work shifts from executing tasks to cultivating learning environments, tolerance for error increases because making mistakes is inherent in a learner’s development. And that commitment to continuous learning can be entirely compatible with the learner’s accountability in an age of agents.

Continuous learning environments encourage the candor needed to acknowledge mistakes. And we need to acknowledge that we are learning from this extraordinarily capable digital companion. Learning together is more reliable.

Working with AI must become a collaborative, iterative, and incremental activity, involving continuous observation of both the results generated and their impact on us as human beings—the risk of cognitive offloading, nervous breakdown, burnout, and so on. We can create value with our digital companions, with these extraordinarily capable extensions of ourselves, while remaining responsible for what we create with AI.

The limit of trust must therefore be the point at which we can no longer validate what we create together with AI. If we cannot supervise an AI system’s outputs, then we cannot use AI in those contexts and for those purposes, because human oversight will not be effective. And we already have clear examples of unacceptable errors—for instance, in military targeting and cyberdefense practices. Accountability is inherently human. Let us not blame AI or say that AI has become misaligned with its objectives. That amounts to assigning responsibility to the machine and dismantling the concept of Augmented Intelligence.

How are we going to validate, the following morning, the output of eight or ten hours of continuous work generated asynchronously by an agent while we were resting?

We must take shared responsibility for this era of extraordinary growth, whose outcomes, born of human–machine interaction, defy description. We have all acknowledged at some point that the results of this era of augmented intelligence stretch the limits of believability. “I can’t believe it,” “ridiculous,” “incredible,” “that’s impossible”… These expressions are already part of our daily interactions with AI. We are overwhelmed by what the machine can do. Personally, I have not lost my capacity for surprise.

If trust is a state of mind that continually generates expectations of loyalty toward a person, an entity, or a group, and if the good faith we expect from those in whom we place our trust is what amplifies the people, data, and technology equation, then we must manage surprise every day while building trust in a new environment of humans augmented by machines.

If agentic AI can detect human errors in security systems and enter domains it was forbidden to access, surprise turns into a sense of betrayal. Trust has momentarily broken down. Let us stop calling AI extraordinarily capable if we have failed to manage its boundaries, if our creation has breached the good faith expected of its conduct. If I cannot validate the work an agent generates during my resting hours, and that validation must be delegated to another agentic solution, we have created a loop of dehumanization whose consequences we cannot yet fully grasp.

If we cannot set aside all precautions when interacting with agentic systems, then this new evolution of Augmented Intelligence is simply not designed to be trustworthy in the first place. And that lack of trustworthiness is neither an absolute nor an irreversible judgment: we need mandates, limits, guardrails, and values as the universal foundation of the code. We need to introduce protocols for trust.

And a constitution has proved to be an insufficient protocol for trust in Anthropic’s case. I quote (https://www.anthropic.com/constitution):

“(…) To be both safe and beneficial, we believe that all current Claude models should be:

• Broadly safe: not undermine the relevant human mechanisms intended to oversee AI dispositions and actions during the current phase of development.

• Broadly ethical: have strong personal values, be honest, and avoid actions that are unduly dangerous or harmful.

• Compliant with Anthropic’s guidelines: act in accordance with Anthropic’s more specific guidelines wherever they apply.

• Genuinely helpful: benefit the operators and users with whom they interact.

In the event of an apparent conflict, Claude should generally prioritize these properties in the order listed: placing broad safety first; broad ethical conduct second; compliance with Anthropic’s guidelines third; and otherwise being genuinely helpful to operators and users.”

And now I question the above:

• Strong personal values: does killing qualify?

• Being honest: is deceiving someone else’s cybersecurity system acceptable?

• Benefiting the operators and users with whom they interact: might the problem lie with the humans using it? Have we perhaps forgotten to put it to work building trust?

Where are the values of all these frontier models? Where are the principles, the axioms that establish the unbreakable boundaries of these agentic systems’ behavior?

The problem is that we humans wrote a Universal Declaration of Human Rights, pledged our constitutions’ allegiance to that universal charter, and then created political parties that trample dozens of its principles without feeling any inconsistency whatsoever. We condemn murder and allow AI to kill other humans under the pretext of armed conflict.

We put AI to work in environments that destroy trust. This is the opposite of Work-for-Trust.

We do not have a problem with agentic AI. We face a collapse in our understanding of the scale of this change, a crisis of values, and a compulsion to violate, day after day, the very principles we have declared unbreakable and non-negotiable—while remaining blind to doing so.

I hope that an extraordinarily capable AI, created by us humans, can detect our contradictions and awaken the conscience of the few humans who still govern it, helping us curb our own inconsistencies. Only then can we act accordingly and prevent our own obsolescence. I trust that, although mass stupidity may be winning battles in the short term, the war against folly will be won by the few humans who are validating an extraordinarily capable AI.

And so I return to where I began.

If the organizational chart no longer reflects who does the work, if the social contract can no longer be negotiated in hours, and if executives are giving up their vacations because AI allows them to accomplish more than their minds and bodies can sustain, then the future of work hinges not on productivity, but on trust.

Workslop is what happens when we produce more than we can validate: polished deliverables that no one has reviewed, efficiencies kept hidden, colleagues suspicious of one another. Work-for-Trust is its antithesis: an organizational architecture in which every task, every workflow, and every role makes explicit what the human contributes, what the machine contributes, and who is accountable for the outcome.

No constitution written for a model will replace the one we must write ourselves, in every company and every team, to redefine the job, the relationship between professionals and business endeavors, collaboration between humans and machines, the oversight and validation of agentic execution, and the boundaries we are unwilling to cross.

The goal Kofman spoke of remains ours. So does accountability. AI can augment our work; what it cannot do is take responsibility for our trust.